StyleGAN-based data enhancement and retinal image blood vessel processing optimization method
By using StyleGAN data enhancement and the IterNet model to optimize retinal image vascular processing, the accuracy problem of retinal image segmentation and classification was solved, high-precision vascular segmentation and classification were achieved, and the model performance and diagnostic assistance effect were improved.
Patent Information
- Application Number
- CN202510709463.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-29
- Publication Date
- 2025-09-16
AI Technical Summary
Existing retinal image segmentation and classification methods have difficulty in accurately extracting blood vessels, especially when dealing with microvessels and diseased blood vessels, which are prone to missed or misdetected. Classification methods are also prone to errors in identifying blood vessel types, affecting the accuracy of disease diagnosis.
StyleGAN data augmentation technology is used to generate high-quality retinal images, and the IterNet and FullSizeUNetBlock models are combined for vascular segmentation and classification. The model is optimized through adaptive instance normalization and progressive training strategy, and vascular structure information is used for post-processing optimization to improve segmentation and classification accuracy.
It improves the accuracy of blood vessel segmentation and classification, provides high-quality data enhancement effects, enhances model performance, assists doctors in accurate diagnosis, and has broad application prospects in medical image analysis.
Smart Images
Figure CN120655523A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of vascular image processing, and in particular to a method for optimizing retinal image vascular processing based on StyleGAN data enhancement. Background Art
[0002] Retinal images can be used to study and diagnose various eye diseases, such as diabetic retinopathy (DR), age-related macular disease (AMD), and glaucoma. Similar to other types of medical images, retinal images exhibit high complexity and great diversity.
[0003] Retinal images are complex structures, containing a wealth of detailed information and diverse vascular morphologies, often accompanied by noise interference. Traditional segmentation methods, whether based on manual feature algorithms or some early deep learning models, struggle to accurately and completely extract blood vessels from complex backgrounds. This is especially true when dealing with microvessels and vessels with morphological changes caused by pathological changes, which can easily lead to missed or false detections. Traditional methods also struggle to accurately segment intersecting or overlapping vascular regions, affecting the continuity and integrity of the vessels in the segmentation results and making it impossible to provide accurate vascular structure information for subsequent diagnosis.
[0004] Existing artificial intelligence (AI)-based computer-aided diagnosis (CAD) systems also face challenges in classification. Because the arteries and veins of retinal blood vessels share a certain similarity in appearance, relying solely on limited features for classification can easily lead to misclassification. For example, when blood vessels are affected by pathological changes, such as color and thickness, existing classification methods may be unable to accurately identify their type, leading to misclassification and compromising doctors' accurate diagnosis of the disease. This challenge urgently needs to be addressed. Summary of the Invention
[0005] To avoid and overcome the technical problems existing in the prior art, the present invention provides a method for data enhancement and retinal image vascular processing optimization based on StyleGAN, which can effectively improve the accuracy of vascular classification.
[0006] To achieve the above object, the present invention provides the following technical solutions:
[0007] A method for data enhancement and retinal image vascular processing optimization based on StyleGAN includes the following steps:
[0008] S1. Obtain the original retinal image and input it into StyleGAN to generate a synthetic retinal image, and then store both the original retinal image and the synthetic retinal image in the dataset;
[0009] S2. Use IterNet to perform blood vessel segmentation and extraction on each retinal image in the dataset to obtain a refined blood vessel map corresponding to each retinal image;
[0010] S3. performing a masking operation on the refined vascular map to generate a corresponding masked vascular map, and merging the masked vascular map with the refined vascular map to obtain a merged vascular map;
[0011] S4. Inputting the merged vascular image into the FullSizeUNetBlock model for training to obtain a vascular category prediction model;
[0012] S5. Optimizing the blood vessel category prediction model to achieve a set prediction accuracy;
[0013] S6. Process the new original retinal image according to steps S2 to S5 to obtain the category of blood vessels in the original retinal image.
[0014] As a further solution of the present invention: the specific content of step S5 is as follows:
[0015] S51. Convert the refined vascular image V into a binary image P by threshold segmentation. The threshold segmentation is expressed as follows:
[0016] p k =1,v k >θ;
[0017] Where p k represents the binary value of the kth pixel in the refined vascular image V; v k represents the probability value of the kth pixel in the refined vascular image V belonging to a vascular pixel; θ represents the predefined threshold;
[0018] S52, defining the vascular branch points, vascular intersections, vascular starting points, and vascular end points in the refined vascular map V as key points;
[0019] S53, connecting each pixel point between adjacent key points in series to form an independent blood vessel segment, each blood vessel segment constitutes a blood vessel set S, S = {S1, ..., S i ,…,s N}, S1 represents the first vessel segment in the refined vessel graph V; S i represents the i-th vessel segment in the refined vessel graph V; S N represents the Nth vessel segment in the refined vessel graph V;
[0020] S54. Calculate the confidence level of each blood vessel segment as a vein or artery. The confidence level calculation formula is as follows:
[0021]
[0022] Where c li Indicates the blood vessel segment S in the value interval l = {artery, vein} i The sum of the confidence differences is used to measure whether the vessel segment is more likely to belong to an artery or a vein; artery,h Indicates the blood vessel segment S i The probability that the h-th pixel in is predicted as an artery by the vessel category prediction model; vein,h Indicates the blood vessel segment S i The probability that the h-th pixel in is predicted as a vein by the blood vessel category prediction model;
[0023] S55, if c li >0, then the vessel segment S i The opposite is an artery, and the opposite is a vein.
[0024] As a further solution of the present invention: after the confidence level is calculated, it can also be updated using the following formula:
[0025] c′ li =c li +∈ ij c lj ;
[0026] Where c' li Represents the confidence level c li Updated confidence value; c lj Indicates the blood vessel segment S in the value interval l = {artery, vein} j The sum of the confidence differences is used to measure whether the vessel segment is more likely to belong to an artery or a vein; ij Indicates the blood vessel segment S i and the adjacent blood vessel segment S connected to it j The influence coefficient between
[0027] ∈ ij =A ij L ij T ij D ij ;
[0028] Where A ij Indicates the blood vessel segment S i With vascular segment S j Unit tangent vector u at a key point i and u j The angle function value between ij Handles possible missing connections between segments; T ij Consider the thickness difference of the segment; D ij Reflects the distance between segments;
[0029] A ij =FA (|α(u i ,u j )-180|);
[0030]
[0031] In the formula, α(u i ,u j ) represents the blood vessel segment S i With vascular segment S j Unit tangent vector u at a key point i and u j The angle between A (x) represents the angle calculation function, x represents the independent variable; m A Indicates a predefined maximum value;
[0032] L ij =F L (α(u i ,w ij ));
[0033] Where w ij From the vascular segment S i The key point to the vessel segment S j The unit vector of α(u i ,w ij ) represents the vector u i and w ij The angle between L (α(u i ,w ij )) represents the relationship between α(u i ,w ij ) is normalized; the thickness difference is calculated as follows:
[0034] T ij =F T (β(S i ,S j ));
[0035] Where, β(S i ,S j ) represents the blood vessel segment S i The average thickness and vessel segment S j The difference between the average thickness of
[0036] D ij =F D (‖p i0 -p j0 ||);
[0037] Where p i0 Indicates the blood vessel segment S iThe starting point coordinates of p j0 Indicates the blood vessel segment S j The starting point coordinates of ||p i0 -p j0 || means p i0 With p j0 The Euclidean distance between D (||p i0 -p j0 ||) to ||p i0 -p j0 ||Normalize. Specific: D ij Represents a measure for line segment S i and S j The distance relationship between variables, whose values will affect S j To S i The degree of label confidence update. If S i and S j Distance is far, D ij will give a small value so that S j To S i The impact of F D Used to normalize the input value, i0 -p j0 ||Map to the interval [0,1] so that different distance values can be calculated on a unified scale to measure the influence of distance on the relationship between line segments. ||p i0 -p j0 || represents line segment S i and S j Starting point (here p i0 Indicates S i The starting point, p j0 Indicates S j By calculating this distance, we can reflect the spatial distance between the two line segments in the image. The larger the distance, the farther the two line segments are apart.
[0038] As a further solution of the present invention, an adaptive instance normalization operation is used in the process of generating the synthetic retina. The adaptive instance normalization operation is specifically expressed as follows:
[0039]
[0040] Where AdaIN(x,y) represents the adaptive instance normalization operation; x represents the feature map extracted from the original retinal image; y represents the style vector; μ(x) represents the pixel mean of the feature map; σ(x) represents the pixel standard deviation of the feature map; β y represents the offset factor learned from the style vector y; γ yrepresents the scaling factor learned from the style vector y.
[0041] As a further solution of the present invention: in the process of using StyleGAN for data enhancement, its training mode adopts a progressive growth strategy, starting from a set low resolution and gradually increasing to the target high resolution according to the set step size, and increasing the resolution after fully training the generator and discriminator at each resolution stage; at the same time, a random noise map is injected before each convolutional layer.
[0042] As a further solution of the present invention: in the process of using StyleGAN for data enhancement, the loss function used is the WGAN-GP loss function.
[0043] Compared with the prior art, the present invention has the following beneficial effects:
[0044] 1. Data augmentation advantage: Using StyleGAN to generate high-quality retinal images, expand the data set, solve traditional data augmentation problems, provide high-quality data for model training, and improve model performance.
[0045] 2. Accurate segmentation and classification: The IterNet structure of the SeqNet model optimizes vascular segmentation. Combined with the UNet classification component and sequential processing, it improves segmentation and classification accuracy, assisting doctors in making accurate diagnoses.
[0046] 3. Optimize classification structure: Post-processing technology unifies labels and propagates predictions based on vascular structure information, optimizing the global structure of classification results to facilitate macroscopic analysis by doctors.
[0047] 4. Good general scalability: The technology is highly versatile and can be expanded to other medical image analysis fields, with broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS
[0048] Figure 1 It is the overall flow chart of the present invention.
[0049] Figure 2 This is the StyleGAN data enhancement flow chart in the present invention.
[0050] Figure 3 This is the SeqNet segmentation and classification network architecture in the present invention. DETAILED DESCRIPTION
[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.
[0052] See also Figures 1 to 3 In an embodiment of the present invention, a method for data enhancement and retinal image blood vessel processing optimization based on StyleGAN includes the following contents:
[0053] 1. Data enhancement.
[0054] A generative adversarial network based on StyleGAN is used to synthesize high-resolution retinal images, expand the training dataset, and improve the imbalance of data distribution.
[0055] StyleGAN is based on the GAN framework and consists of a generator and a discriminator. The generator receives random noise and generates images; the discriminator determines the authenticity of the image. The two continuously improve their performance through adversarial training. See the flowchart for details. Figure 2 The key technologies for model training are as follows:
[0056] First, the input random noise z is mapped to a style vector w through a multi-layer perceptron. This vector contains rich style information. During the generation process, the adaptive instance normalization (AdaIN) operation is used. Its formula is AdaIN(x,y) represents the adaptive instance normalization operation; x represents the feature map extracted from the original retinal image; y represents the style vector; μ(x) represents the pixel mean of the feature map; σ(x) represents the pixel standard deviation of the feature map; β y represents the offset factor learned from the style vector y; γ y Represents the scaling factor learned from the style vector y; adjusts the mean and standard deviation of the feature map according to the style vector to achieve style modulation, thereby flexibly controlling the image style.
[0057] Training uses a progressive growth strategy, starting with a low resolution (e.g., 4×4) and gradually increasing to the target high resolution (e.g., 1024×1024). After fully training the generator and discriminator at each resolution stage, the resolution is increased. This reduces the difficulty and computational complexity of training and avoids the instability issues associated with high-resolution training. Furthermore, a random noise pattern is injected before each convolutional layer to control image detail and texture, making the images appear more natural and realistic.
[0058] The loss function uses WGAN-GP loss. The discriminator maximizes the Wasserstein distance between the real and generated images and satisfies the gradient penalty, while the generator minimizes the distance to ensure training stability and image quality.
[0059] In addition, StyleGAN supports style mixing, which can use different style vectors in different convolutional layers to generate images that combine multiple style features.
[0060] 2. Retinal Vessel Segmentation
[0061] The improved UNet model (IterNet) is used to perform blood vessel segmentation on the enhanced retinal image and extract the blood vessel area.
[0062] In order to accurately segment the blood vessels, the structural redundancy in the vascular system is utilized to avoid interference from inaccurate segmentation at the edge of the optical disc. The IterNet structure is used in this part. Its core idea is to shift the focus of the deep learning model from processing each pixel in the original input image to the entire vascular network system. More specifically, the inaccurate blood vessel segmentation results are improved to more accurate blood vessel segmentation results, rather than directly mapping the original input image to an accurate segmentation result. The IterNet architecture mainly consists of a UNet and multiple iterative mini-UNets. This design achieves accurate segmentation of retinal blood vessels as follows:
[0063] Foundational Module (UNet): IterNet uses UNet as its foundational module due to its outstanding performance in various segmentation tasks, particularly in medical applications. UNet performs preliminary processing on the raw retinal image and outputs a single-channel vascular probability map as input to subsequent modules. This module plays a key role in extracting basic vascular features from the raw image, laying the foundation for subsequent accurate segmentation.
[0064] Refinement module (mini-UNet): The architecture of the refinement module is mini-UNet, and its input is the 32-channel feature map output by the penultimate layer of the previous module. Compared with the single-channel vascular probability map, this feature map contains richer information. Mini-UNet is a lightweight UNet architecture with a small number of parameters because its input feature map is simpler than the original retinal image containing background and noise. In IterNet, all mini-UNets share the same weights and biases. The input of the first mini-UNet comes from the feature map of the penultimate layer of the basic module, and the remaining mini-UNets follow this process. This allows the same module to process different inputs in a single forward propagation, continuously expose new patterns of vascular segmentation failures, and thus learn optimization from a small number of training samples.
[0065] Skip-connections: To solve the problem of difficulty in obtaining input information for the upper layers in multi-layer networks and the gradient vanishing problem that is prone to occur in deep learning models, IterNet adds skip connections similar to UNet, which specifically include three types: one is the intra-module connection, which is used to connect the encoding layer and decoding layer of each module; the second is the connection from the basic UNet to all refined mini-UNets, so that the features of the first layer of the basic UNet (close to the input retinal image) can be spliced with the features of the first layer of each mini-UNet, providing reference information of the original image for the upper-layer modules; the third is the connection between mini-UNets, which is inspired by the dense connections of dense networks and splices the features of the lower-layer modules with the features of the upper-layer modules.
[0066] 3. Artery / Veneous Classification
[0067] The classification part of UNet is used to classify the segmented blood vessels into arteries / veins, and the classification accuracy is improved through sequential processing (segmentation first and then classification).
[0068] After the input retinal image x is processed by IterNet to obtain the refined vascular map v, the background pixels of the input image x are masked using the formula x'=x⊙v. That is, the refined vascular map v obtained is not a complete image, but a cut-out blood vessel. Element-wise multiplication with the original image can set the background pixels except the blood vessel pixels to black, reducing the complexity of the input image and allowing the classification flow to focus on the differences between blood vessels. Using the full-size UNet block to process x', the output of the classification flow will be merged with the segmentation result, and the probability o of each pixel belonging to the background, artery or vein will be obtained through softmax. l During training, IterNet uses segmentation labels to calculate cross entropy loss, and the entire model is jointly trained with a batch size of 16. The optimizer uses Adam and the learning rate is set to 0.001.
[0069] 4. Post-processing optimization
[0070] By unifying intra-segment labels and propagating inter-segment predictions, the classification results are further optimized and misclassification caused by local features is reduced.
[0071] This scheme improves accuracy through post-processing intra-segment label unification and inter-segment prediction propagation technology.
[0072] (1) Unification of intra-segment labels: First, the binary image p of the detected blood vessels is generated from the output v of SeqNet by the following formula: k =1{v k >θ}, where p k represents the binary value of the kth pixel in the refined vascular image V; v krepresents the probability that the kth pixel in the refined vessel map V belongs to a vessel pixel; θ represents a predefined threshold. All keypoints are detected, including intersections between vessels and endpoints (i.e., starting and ending points) of vessels. Intersections are detected by finding vessel pixels with more than two neighbors on the skeleton image, while endpoints have no more than one neighbor. Skeleton pixels between connected keypoints are extracted as a segment.
[0073] Let S = {S i |i=1,…,N} is the set of all N segments extracted from P, S1 represents the first vessel segment in the refined vessel graph V; S i represents the i-th vessel segment in the refined vessel graph V; S N Represents the Nth vessel segment in the refined vessel map V. The segment S is calculated by the following formula i Confidence c that l = {artery, vein} li : O artery,h Indicates the blood vessel segment S i The probability that the h-th pixel in is predicted as an artery by the vessel category prediction model; vein,h Indicates the blood vessel segment S i The probability that the h-th pixel in is predicted as a vein by the vessel category prediction model.
[0074] The actual prediction can be obtained by comparing c li To complete, that is, if the confidence is greater than 0, then S i It is an artery, otherwise it is a vein.
[0075] (2) Inter-segment prediction propagation: To address errors near intersections and branch points, an inter-segment prediction propagation method is introduced, where the label of a segment is propagated to its connected segments. This is based on the observation that classification failures are often accompanied by low label confidence, and these errors can be corrected by the influence of connected high-confidence segments. The propagation should be based on the similarity between connected segments in terms of shape, direction, etc. If two segments have similar shapes, close locations, and similar flow directions, then they are likely to belong to the same vessel.
[0076] Based on this observation, after the confidence level is calculated, it can be updated using the following formula:
[0077] c′ li =C li +∈ ij C lj ;
[0078] Where c' li Represents the confidence c li Updated confidence value; c lj Indicates the blood vessel segment S in the value interval l = {artery, vein}j The sum of the confidence differences is used to measure whether the vessel segment is more likely to belong to an artery or a vein; ij Indicates the blood vessel segment S i and the adjacent blood vessel segment S connected to it j The influence coefficient between
[0079] 5. The method for data enhancement and retinal image vascular processing optimization based on StyleGAN according to claim 4, wherein the influence coefficient is calculated as follows:
[0080] ∈ ij =A ij L ij T ij D ij ;
[0081] Where A ij Indicates the blood vessel segment S i With vascular segment S j The unit tangent vector u at the intersection point i and u j The angle function value between ij Indicates the blood vessel segment S i With vascular segment S j The missing connection between T ij Indicates the blood vessel segment S i With vascular segment S j Thickness difference between ij Indicates the blood vessel segment S i With vascular segment S j The distance between them.
[0082] 6. The method for data enhancement and retinal image vascular processing optimization based on StyleGAN according to claim 5, wherein the angle function value is calculated as follows:
[0083] A ij =F A (|α(u i ,u j )-180|);
[0084]
[0085] In the formula, α(u i ,u j ) represents the blood vessel segment S i With vascular segment S j The unit tangent vector u at the intersection point i and u j The angle between |α(u i ,u j)-180|that is, two line segments u i and u j Angle α(u i ,u j ) and 180 degrees; F A (x) represents the normalization function, x represents the independent variable; m A Indicates the predefined maximum value; F T 、F D With F A The definition is similar.
[0086] The missing connections are calculated as follows:
[0087] L ij =F L (α(u i ,w ij ));
[0088] Where w ij From the vascular segment S i From the starting point to the vascular segment S j The unit vector between the starting points;
[0089] α(u i ,w ij ) represents the vector u i and w ij The angle between L (α(u i ,w ij )) represents the relationship between α(u i ,w ij ) for normalization;
[0090] The thickness difference is calculated as follows:
[0091] T ij =F T (β(S i ,S j ));
[0092] Where, β(S i ,S j ) represents the blood vessel segment S i The average thickness and vessel segment S j The difference between the average thickness of
[0093] The distance calculation formula is as follows:
[0094] D ij =F D (||p i0 -p j0 ||);
[0095] Where pi0 Indicates the blood vessel segment S i The starting point coordinates of p j0 Indicates the blood vessel segment S j The starting point coordinates of ||p i0 -p j0 || means p i0 With p j0 The Euclidean distance between D (||p i0 -p j0 ||) to ||p i0 -p j0 ||Perform normalization.
[0096] Apply this update rule to all extracted segments. After several iterations, the predicted labels are corrected. Note that a segment has two endpoints, and A ij , L ij and D ij In each segment S i and S j Only one endpoint is involved in each case. Confidence is updated for all four endpoint combinations.
[0097] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.
Claims
1. A method for data enhancement and retinal image vascular processing optimization based on StyleGAN, characterized in that: The following steps are involved: S1. Obtain the original retinal image and input it into StyleGAN to generate a synthetic retinal image, and then store both the original retinal image and the synthetic retinal image in the dataset; S2. Use IterNet to perform blood vessel segmentation and extraction on each retinal image in the dataset to obtain a refined blood vessel map corresponding to each retinal image; S3. performing a masking operation on the refined vascular map to generate a corresponding masked vascular map, and merging the masked vascular map with the refined vascular map to obtain a merged vascular map; S4. Inputting the merged vascular image into the FullSizeUNetBlock model for training to obtain a vascular category prediction model; S5. Optimizing the blood vessel category prediction model to achieve a set prediction accuracy; S6. Process the new original retinal image according to steps S2 to S5 to obtain the category of blood vessels in the original retinal image.
2. The method for data enhancement and retinal image vascular processing optimization based on StyleGAN according to claim 1, characterized in that: The specific content of step S5 is as follows: S51. Convert the refined vascular image V into a binary image P by threshold segmentation. The threshold segmentation is expressed as follows: p k =1,v k >θ; Where p k represents the binary value of the kth pixel in the refined vascular image V; v k represents the probability value of the kth pixel in the refined vascular image V belonging to a vascular pixel; θ represents the predefined threshold; S52, defining the vascular branch points, vascular intersections, vascular starting points, and vascular end points in the refined vascular map V as key points; S53, connecting each pixel point between adjacent key points in series to form an independent blood vessel segment, each blood vessel segment constitutes a blood vessel set S, S = {S1, ..., S i ,…,S N }, S1 represents the first vessel segment in the refined vessel graph V; S i represents the i-th vessel segment in the refined vessel graph V; S N represents the Nth vessel segment in the refined vessel graph V; S54, calculating the confidence level of each blood vessel segment as belonging to a vein or an artery; S55. If the confidence level is greater than 0, then the vessel segment S i The opposite is an artery, and the opposite is a vein.
3. The method for data enhancement and retinal image vascular processing optimization based on StyleGAN according to claim 2, characterized in that: The confidence calculation formula is as follows: Where c li Indicates that in the value interval l = artery, vein, blood vessel segment S i The sum of the confidence differences is used to measure whether the vessel segment is more likely to belong to an artery or a vein; artery,h Indicates the blood vessel segment S i The probability that the h-th pixel in is predicted as an artery by the vessel category prediction model; vein,h Indicates the blood vessel segment S i The probability that the h-th pixel in is predicted as a vein by the vessel category prediction model.
4. The method for data enhancement and retinal image vascular processing optimization based on StyleGAN according to claim 3, characterized in that: After the confidence level is calculated, it can be updated using the following formula: c' li =c li +∈ ij C lj ; Where c′ li Represents the confidence level c li Updated confidence value; c lj Indicates the blood vessel segment S in the value interval l = {artery, vein} j The sum of the confidence differences is used to measure whether the vessel segment is more likely to belong to an artery or a vein; ij Indicates the blood vessel segment S i and the adjacent blood vessel segment S connected to it j The influence coefficient between 5. The method for data enhancement and retinal image blood vessel processing optimization based on StyleGAN according to claim 4, characterized in that: The calculation formula of the influence coefficient is as follows: ∈ ij =A ij L ij T ij D ij ; Where A ij Indicates the blood vessel segment S i With vascular segment S j The unit tangent vector u at the intersection point i and u j The angle function value between ij Indicates the blood vessel segment S i With vascular segment S j The missing connection between T ij Indicates the blood vessel segment S i With vascular segment S j Thickness difference between ij Indicates the blood vessel segment S i With vascular segment S j The distance between them.
6. The method for data enhancement and retinal image blood vessel processing optimization based on StyleGAN according to claim 5, characterized in that: The calculation formula of the angle function value is as follows: A ij =F A (|α(u i ,u j )-180|); In the formula, α(u i ,u j ) represents the blood vessel segment S i With vascular segment S j The unit tangent vector u at the intersection point i and u j The angle between |α(u i ,u j )-180|that is, two line segments u i and u j Angle α(u i ,u j ) and 180 degrees; F A (x) represents the normalization function, x represents the independent variable; m A Indicates a predefined maximum value; The missing connections are calculated as follows: L ij =F L (a(u i ,w ij )); Where w ij From the vascular segment S i From the starting point to the vascular segment S j The unit vector between the starting points; α(u i ,w ij ) represents the vector u i and w ij The angle between L (α(u i ,w ij )) represents the relationship between α(u i ,w ij ) for normalization; The thickness difference is calculated as follows: T ij =F T (β(S i ,S j )); Where, β(S i ,S j ) represents the blood vessel segment S i The average thickness and vessel segment S j The difference between the average thickness of The distance calculation formula is as follows: D ij =F D (||p i0 -p j0 ||); Where p i0 Indicates the blood vessel segment S i The starting point coordinates of p j0 Indicates the blood vessel segment S j The starting point coordinates of ||p i0 -p j0 || means p i0 With p j0 The Euclidean distance between D (||p i0 -p j0 ||) to ||p i0 -p j0 ||Perform normalization.
7. A method for data enhancement and retinal image vascular processing optimization based on StyleGAN according to any one of claims 1 to 6, characterized in that: In the process of synthetic retina generation, adaptive instance normalization operation is used; the specific expression of adaptive instance normalization operation is as follows: Where AdaIN(x,y) represents the adaptive instance normalization operation; x represents the feature map extracted from the original retinal image; y represents the style vector; μ(x) represents the pixel mean of the feature map; σ(x) represents the pixel standard deviation of the feature map; β y represents the offset factor learned from the style vector y; γ y represents the scaling factor learned from the style vector y.
8. The method for data enhancement and retinal image blood vessel processing optimization based on StyleGAN according to claim 7, characterized in that: In the process of using StyleGAN for data augmentation, its training mode adopts a progressive growth strategy, starting from a set low resolution and gradually increasing to the target high resolution according to the set step size. The resolution is increased only after the generator and discriminator are fully trained at each resolution stage; at the same time, a random noise map is injected before each convolutional layer.
9. The method for data enhancement and retinal image blood vessel processing optimization based on StyleGAN according to claim 8, characterized in that: In the process of using StyleGAN for data enhancement, the loss function used is the WGAN-GP loss function.